Can we reliably detect whether a text was written by a human or an AI? This question has become increasingly crucial with the rise of powerful language models like ChatGPT. New research suggests a surprisingly simple yet effective method: checking for grammar errors. A paper titled "Who Wrote This? The Key to Zero-Shot LLM-Generated Text Detection Is GECScore" introduces a novel approach called GECScore. The core idea is that humans, even skilled writers, make more grammatical mistakes than large language models (LLMs). LLMs are trained on massive datasets of grammatically correct text, making them less prone to errors. GECScore works by using a grammar error correction (GEC) model. The GEC model corrects the input text, and then a similarity score is calculated between the original and corrected versions. If the similarity is high, the text is likely LLM-generated. If the similarity is low, suggesting more corrections were needed, it's likely human-written. This method is "zero-shot," meaning it doesn't require training on examples of AI-generated text. It's also "black-box," meaning it doesn't need access to the inner workings of the LLM that generated the text. This makes it incredibly versatile. The researchers tested GECScore against a range of existing detection methods, including those used by OpenAI. Impressively, GECScore outperformed them all, achieving an average accuracy of 98.7%. It even proved robust against attempts to fool it with paraphrased or slightly altered text. While GECScore relies on having a good GEC model, the research opens exciting possibilities. Future work could explore using LLMs *themselves* as the GEC models, potentially leading to even more accurate detection. As AI-generated text becomes more prevalent, tools like GECScore will be essential for maintaining trust and integrity in written content, from academic papers to news articles.